Differential convolutional neural network

Mehmet Sarigul, Buse Melis Özyildirim, Mutlu Avci

Abstract: Highlights•A novel convolution technique named as Differential Convolution is proposed.•The technique aims to transfer directional activation differences to the next layer.•The technique considers how convolved features change on a feature map.•An updated back-propagation algorithm compatible with the technique is developed.